File size: 3,334 Bytes
092488e 5f18130 092488e fb23377 092488e d55cdb1 092488e a2bce82 307eb7b 092488e d55cdb1 f6775a2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 |
---
license: llama3.1
library_name: goodfire-llama-3.1-8b-instruct-sae-l19
language:
- en
tags:
- mechanistic interpretability
- sparse autoencoder
- llama
- llama-3
---
## Model Information
The Goodfire SAE (Sparse Autoencoder) for [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct)
is an interpreter model designed to analyze and understand
the model's internal representations. This SAE model is trained specifically on layer 19 of
Llama 3.1 8B and achieves an L0 count of 91, enabling the decomposition of complex neural activations
into interpretable features. The model is optimized for interpretability tasks and model steering applications,
allowing researchers and developers to gain insights into the model's internal processing and behavior patterns.
As an open-source tool, it serves as a foundation for advancing interpretability research and enhancing control
over large language model operations.
__Model Creator__: [Goodfire](https://huggingface.co/Goodfire), built to work with [Meta's Llama models](https://huggingface.co/meta-llama)
By using __Goodfire/Llama-3.1-8B-Instruct-SAE-l19__ you agree to the [LLAMA 3.1 COMMUNITY LICENSE AGREEMENT](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct/blob/main/LICENSE)
## Intended Use
By open-sourcing SAEs for leading open models, especially large-scale
models like Llama 3.1 8B, we aim to accelerate progress in interpretability research.
Our initial work with these SAEs has revealed promising applications in model steering,
enhancing jailbreaking safeguards, and interpretable classification methods.
We look forward to seeing how the research community builds upon these
foundations and uncovers new applications.
#### Feature labels
To explore the feature labels check out the [Goodfire Ember SDK](https://www.goodfire.ai/blog/announcing-goodfire-ember/),
the first hosted mechanistic interpretability API.
The SDK provides an intuitive interface for interacting with these
features, allowing you to investigate how Llama processes information
and even steer its behavior. You can explore the SDK documentation at [docs.goodfire.ai](https://docs.goodfire.ai).
## How to use
View the notebook guide below to get started.
<a href="https://colab.research.google.com/drive/1IBMQtJqy8JiRk1Q48jDEgTISmtxhlCRL" target="_blank">
<img
src="https://colab.research.google.com/assets/colab-badge.svg"
alt="Open in Colab"
width="200px"
style={{ pointerEvents: "none" }}
/>
</a>
## Training
We trained our SAE on activations harvested from Llama-3.1-8B-Instruct on the [LMSYS-Chat-1M dataset](https://arxiv.org/pdf/2309.11998).
## Responsibility & Safety
Safety is at the core of everything we do at Goodfire. As a public benefit
corporation, we’re dedicated to understanding AI models to enable safer, more reliable
generative AI. You can read more about our comprehensive approach to
safety and responsible development in our detailed [safety overview](https://www.goodfire.ai/blog/our-approach-to-safety/).
Toxic features were removed prior to the release of this SAE. If you are a safety researcher that
would like access to the features we’ve removed, you can reach out at <a href="mailto:[email protected]">[email protected]</a> for access. |